Labs

Abstract: Despite the high performance recently achieved, automatic speech recognition (ASR) systems struggle with the variability of real-world spontaneous speech. Evidence that scaling up training data does not guarantee improvement suggests that an approach grounded in understanding speech phenomena beyond signal processing is needed to optimise ASR models. Based on selected datasets, this hands-on session offers a guided exploration of the key correlates of spoken language, including phonetic-prosodic variation, reduction, disfluency phenomena, the sociolinguistic variability of data often subsumed under the overarching label “spontaneous speech”, and the relevant acoustic features for characterising and distinguishing speech styles. By providing a workflow for modelling human speech behaviour and insight into integrating linguistic knowledge into ASR development, the lab highlights how a deeper understanding of speech variability can inform optimised training data, hybrid modelling approaches, and more robust evaluation strategies, ultimately leading to improved speech models for both research and practical applications.

Abstract: How can we control Large Language Models (LLMs) and adapt them to our needs in an interpretable manner? In this laboratory we will start with the simplest “black box” control techniques, prompting and fine-tuning, and progressively move toward methods that directly interact with model internals in a more interpretable way. We will demonstrate how steering vectors and feature-based interventions derived from sparse representations, such as those obtained via Sparse Autoencoders and Transcoders, enable not only precise control over model behavior but also transparent and interpretable manipulation of its internal activations. Through live coding and interactive exercises, attendees will gain practical insights into the emerging field of mechanistic interpretability, learning how to locate concepts within a model’s internal representations and leverage this knowledge to elicit specific, desired behaviors.

Abstract: This hands-on lab introduces data curation and efficient Speech-to-Text fine-tuning for minority languages, using South Tyrolean dialects as a running example. Starting from raw recordings, participants will be introduced to a typical audio data pipline, transitioning to clean datasets using specialized scripts for automated conversion and chunking. We discuss common pitfalls (noise, inconsistent orthography, imbalance, domain shift) and simple quality-control reports to catch them early. We show how the Eurac Research CLARIN Centre supports this workflow, from metadata standards to long-term